Piyush ChandraAI Research Engineer · Ethara.AI
ContactRésumé →
Overview

Piyush Chandra

AI Research Engineer at Ethara.AI · Delhi NCR, India

I work on the reinforcement-learning side of training large language models. I design the environments models train in, build the programs that grade their answers (verifiers), and shape the reward signals that teach them to get better, across reasoning, code, mathematics, and agentic tool-use.

Reinforcement learningLLM post-trainingReward modelingVerifier engineeringAgentic AIEvaluation design
Current role
AI Research Engineer
at Ethara.AI
Time in role
1 yr · 8 mo
since Oct 2024
Focus area
RL post-training
for large language models
Experience

Ethara.AI

Current
Oct 2024 – Present · 1 yr · 8 mo
AI Research Engineer · Gurgaon, India
RLRLVRRLHFReward modelingVerifiersPost-trainingAgentic AI
  • •Drive end-to-end, Python-based reinforcement-learning and data-driven post-training pipelines that capture chain-of-thought, optimize model behavior, and improve task-based accuracy.
  • •Design and build simulated evaluation environments for Agentic AI systems and conversational AI agents, spanning complex multi-turn reasoning, code generation, mathematics, and agentic tool-use.
  • •Engineer programmatic verifiers and rubric-based evaluation frameworks that produce reliable, scalable reward signals for post-training.
  • •Author high-difficulty Python tasks, ground-truth solutions, and rubric-based evaluations used in post-training and model assessment.
  • •Build internal tooling and dashboards (Python, plus React + TypeScript) supporting model training, evaluation workflows, and operational efficiency.
Key achievements
  • ▸Strengthened model performance and reliability by architecting verifiable, reward-based RL environments and scalable evaluation systems.
  • ▸Improved training efficiency and decision-making accuracy through high-quality task design, ground-truth datasets, and rubric-based validation frameworks.
  • ▸Raised workflow productivity and cross-team collaboration by building internal tools that streamline training, evaluation, and experimentation.

Feynn Labs

Past
Jul 2024 – Sep 2024 · 3 months
Data Science Intern · Team Lead · Remote, India
Team leadershipClusteringDecision treesEV market
  • •Led a small project team on a Government of India initiative analysing the domestic Electric Vehicle (EV) market, owning scope and milestones and running end-to-end market-segmentation analysis to identify levers for accelerating consumer adoption.
  • •Built K-Means clustering and decision-tree models on multi-source datasets (sales, demographic, geographic, charging infrastructure) to profile high-propensity consumer cohorts across Indian states.
  • •Mapped supply-side chokepoints across battery sourcing, charging infrastructure, and OEM distribution; engineered the full Python data pipeline (pandas, NumPy, scikit-learn) and presented findings via dashboards and a written report.
Key achievements
  • ▸Identified high-potential consumer segments and supply-side bottlenecks across the EV ecosystem, enabling targeted policy recommendations and go-to-market strategy.
  • ▸Delivered actionable insights through advanced analytics and dashboards, improving decision-making for accelerating EV adoption.

Galgotias University

Education
Aug 2019 – Aug 2023 · 4 yrs
B.Tech, Computer Science with AI/ML (Honors) · Uttar Pradesh, India
Technical Club LeadGDSC Technical Lead
Selected projectsWhat I work on at Ethara.AI
Specific environments, verifiers, and evaluation suites are confidential; descriptions are kept high-level and discussed in more depth on request.
Post-training · RLHF · RLVR · Model training

LLM post-training & model-training pipelines

End-to-end, Python-based post-training of large language models: supervised fine-tuning (SFT), reward modeling, RLHF, and RLVR with policy-optimization methods (GRPO, GTPO) and parameter-efficient LoRA fine-tuning. Pipelines capture chain-of-thought, optimize model behavior, and lift task-based accuracy.

Tools
PyTorchHuggingFaceSFTReward ModelingGRPOGTPOLoRAW&B
Reasoning · Code · Math · Agentic tool-use

RL environments with verifiable rewards

Reinforcement-learning environments across reasoning, code generation, mathematics, and agentic tool-use. Each environment specifies a task distribution, a verifier program, and a deterministic reward function used for post-training and evaluation of frontier language models.

Tools
PythonPyTorchRLVRDockerpytest
Reasoning · Code · Math

Programmatic verifiers

Verifier programs that produce reliable, mechanically-checkable reward signals across multiple domains. Each verifier is intentionally simpler than the policy it grades, a property that keeps the reward signal trustworthy as the model improves.

Tools
PythonpytestHermetic execution
Post-training · Model assessment

High-difficulty tasks and rubric evaluations

High-difficulty Python tasks with ground-truth solutions and rubric-based evaluations used in post-training and model assessment. Rubrics cover partial credit, reasoning quality, and adherence to specification.

Tools
PythonRubric designLLM-as-judge
Core competenciesStrengths I bring to a team
Model performance optimization & accuracyScalable training-pipeline effectivenessData quality & evaluation reliabilityAutomation of model-training workflowsSystem robustness & stabilityAnalytical, decision-enabling insightProcess standardization & operational consistencyCross-functional execution & delivery alignmentContinuous innovation & performance improvement
Skills & tools
Reinforcement Learning
RL Environment DesignReinforcement Learning with Verifiable Rewards (RLVR)RL from Human Feedback (RLHF)Reward ModelingPolicy OptimizationVerifier EngineeringEvaluation Design
LLM Post-Training
Supervised Fine-Tuning (SFT)Reward ModelingGRPOGTPOLoRA Fine-TuningChain-of-Thought DesignRubric-Based Evaluation
Agentic AI
Agentic AI ArchitecturesAgent ToolsConversational AI AgentsMulti-Turn ReasoningSimulated Evaluation EnvironmentsAgentic Tool-Use Evaluation
AI / Machine Learning
Natural Language ProcessingPrompt EngineeringMultimodal ReasoningClassificationClusteringPredictive Modeling
Languages
PythonC++TypeScriptSQL
Frameworks & Tools
PyTorchHuggingFaceWeights & Biases (W&B)pandasNumPyscikit-learnReactNext.jsJupyterpytestDockerGit
Cloud & Infrastructure
AWSGCPNative API Integration
Data
Exploratory Data Analysis (EDA)Data CurationData CleaningFeature EngineeringData Visualization
Contact

Email is the fastest way to reach me.